Representation of Chemical Kinetics by Artificial Neural Networks for Large Eddy Simulations

Similar documents
Large-Eddy Simulation of Turbulent Flames in Syn-Gas Fuel-Air Mixtures

Large Eddy Simulation of Bluff Body Stabilized Premixed Flame

Large Eddy Simulation of a Swirling Non-Premixed Flame

Insights into Model Assumptions and Road to Model Validation for Turbulent Combustion

LARGE EDDY SIMULATION OF SYNGAS-AIR DIFFUSION FLAMES WITH ARTIFICIAL NEURAL NETWORKS BASED CHEMICAL KINETICS

Lecture 14. Turbulent Combustion. We know what a turbulent flow is, when we see it! it is characterized by disorder, vorticity and mixing.

Reaction Rate Closure for Turbulent Detonation Propagation through CLEM-LES

A comparison between two different Flamelet reduced order manifolds for non-premixed turbulent flames

Approximation of chemical reaction rates in turbulent combustion simulation

D. VEYNANTE. Introduction à la Combustion Turbulente. Dimanche 30 Mai 2010, 09h00 10h30

A G-equation formulation for large-eddy simulation of premixed turbulent combustion

Impact of numerical method on auto-ignition in a temporally evolving mixing layer at various initial conditions

S. Kadowaki, S.H. Kim AND H. Pitsch. 1. Motivation and objectives

ANSYS Advanced Solutions for Gas Turbine Combustion. Gilles Eggenspieler 2011 ANSYS, Inc.

A priori Tabulation of Turbulent Flame Speeds via a Combination of a Stochastic Mixing Model and Flamelet Generated Manifolds 5

Regularization modeling of turbulent mixing; sweeping the scales

Before we consider two canonical turbulent flows we need a general description of turbulence.

DARS overview, IISc Bangalore 18/03/2014

Exercises in Combustion Technology

Flamelet Analysis of Turbulent Combustion

Numerical Simulation of Pollutant Emission and Flame Extinction in Lean Premixed Systems. Gilles Eggenspieler

Towards regime identification and appropriate chemistry tabulation for computation of autoigniting turbulent reacting flows

The lattice Boltzmann equation (LBE) has become an alternative method for solving various fluid dynamic

PDF Modeling and Simulation of Premixed Turbulent Combustion

DNS and LES of Turbulent Combustion

Large-eddy simulations of turbulent reacting stagnation point flows

Large Eddy Simulation of Soot Formation in Turbulent Premixed Flame

PDF modeling and simulation of premixed turbulent combustion

Large Eddy Simulation of bluff body stabilized premixed and partially premixed combustion. Ionuţ Porumbel

Reliability of LES in complex applications

An Introduction to Theories of Turbulence. James Glimm Stony Brook University

Asymptotic Structure of Rich Methane-Air Flames

LES of the Sandia Flame D Using an FPV Combustion Model

A Priori Model for the Effective Lewis Numbers in Premixed Turbulent Flames

DNS of Reacting H 2 /Air Laminar Vortex Rings

Large-eddy simulation of an industrial furnace with a cross-flow-jet combustion system

A dynamic global-coefficient subgrid-scale eddy-viscosity model for large-eddy simulation in complex geometries

fluctuations based on the resolved mean flow

Budget analysis and model-assessment of the flamelet-formulation: Application to a reacting jet-in-cross-flow

2D Direct Numerical Simulation of methane/air turbulent premixed flames under high turbulence intensity Julien Savre 04/13/2011

CFD Analysis of Vented Lean Hydrogen Deflagrations in an ISO Container

Predicting NO Formation with Flamelet Generated Manifolds

Survey of Turbulent Combustion Models for Large Eddy Simulations of Propulsive Flowfields

Modeling flame brush thickness in premixed turbulent combustion

Best Practice Guidelines for Combustion Modeling. Raphael David A. Bacchi, ESSS

Effects of Damköhler number on flame extinction and reignition in turbulent nonpremixed flames using DNS

REDIM reduced modeling of quenching at a cold inert wall with detailed transport and different mechanisms

FLAME AND EDDY STRUCTURES IN HYDROGEN AIR TURBULENT JET PREMIXED FLAME

Process Chemistry Toolbox - Mixing

DARS Digital Analysis of Reactive Systems

Consistent turbulence modeling in a hybrid LES/RANS PDF method for non-premixed flames

CHAPTER 7 SEVERAL FORMS OF THE EQUATIONS OF MOTION

LES Approaches to Combustion

Turbulent Premixed Combustion

Extinction Limits of Premixed Combustion Assisted by Catalytic Reaction in a Stagnation-Point Flow

LARGE EDDY SIMULATION OF MASS TRANSFER ACROSS AN AIR-WATER INTERFACE AT HIGH SCHMIDT NUMBERS

Multiscale Computation of Isotropic Homogeneous Turbulent Flow

Numerical Methods in Aerodynamics. Turbulence Modeling. Lecture 5: Turbulence modeling

A validation study of the flamelet approach s ability to predict flame structure when fluid mechanics are fully resolved

Overview of Turbulent Reacting Flows

LES of an auto-igniting C 2 H 4 flame DNS

Simulation of Turbulent Lifted Flames and their Transient Propagation

Colloquium FLUID DYNAMICS 2012 Institute of Thermomechanics AS CR, v.v.i., Prague, October 24-26, 2012 p.

Introduction to Turbulence and Turbulence Modeling

Direct numerical simulation of a turbulent reacting jet

ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD

7. Basics of Turbulent Flow Figure 1.

Laminar Premixed Flames: Flame Structure

IMPROVED POLLUTANT PREDICTIONS IN LARGE-EDDY SIMULATIONS OF TURBULENT NON-PREMIXED COMBUSTION BY CONSIDERING SCALAR DISSIPATION RATE FLUCTUATIONS

Coupling tabulated chemistry with large-eddy simulation of turbulent reactive flows

Numerical Investigation of Ignition Delay in Methane-Air Mixtures using Conditional Moment Closure

Thermoacoustic Instabilities Research

Turbulence Modeling I!

Application of FGM to DNS of premixed turbulent spherical flames

Turbulent Boundary Layers & Turbulence Models. Lecture 09

CFD modeling of combustion

Turbulence: Basic Physics and Engineering Modeling

TOPICAL PROBLEMS OF FLUID MECHANICS 97

Mass Transfer in Turbulent Flow

Flow and added small-scale topologies in a turbulent premixed flame

NUMERICAL ANALYSIS OF TURBULENT FLAME IN AN ENCLOSED CHAMBER

Computational Fluid Dynamics 2

Multi-Scale Modeling of Turbulence and Microphysics in Clouds. Steven K. Krueger University of Utah

Turbulence is a ubiquitous phenomenon in environmental fluid mechanics that dramatically affects flow structure and mixing.

The influence of C ϕ is examined by performing calculations with the values C ϕ =1.2, 1.5, 2.0 and 3.0 for different chemistry mechanisms.

ADVANCED DES SIMULATIONS OF OXY-GAS BURNER LOCATED INTO MODEL OF REAL MELTING CHAMBER

Modeling and Simulation of Plasma-Assisted Ignition and Combustion

Large eddy simulation of turbulent flow over a backward-facing step: effect of inflow conditions

Lecture 4: Perceptrons and Multilayer Perceptrons

Simulations of freely propagating turbulent premixed flames

Lecture 8 Laminar Diffusion Flames: Diffusion Flamelet Theory

Massimo GERMANO Politecnico di Torino

Lecture 10 Turbulent Combustion: The State of the Art

Multi-dimensional transport: DNS analysis and incorporation into the Reaction-Diffusion Manifold (REDIM) method

Address for Correspondence

LES modeling of heat and mass transfer in turbulent recirculated flows E. Baake 1, B. Nacke 1, A. Umbrashko 2, A. Jakovics 2

DNS of auto ignition in turbulent diffusion H 2 /air

Physical Diffusion Cures the Carbuncle Phenomenon

Prediction of unsteady heat transfer from a cylinder in crossflow

Efficient and accurate time-integration of combustion chemical kinetics using artificial neural networks

Transcription:

43rd AIAA/ASME/SAE/ASEE Joint Propulsion Conference & Exhibit 8 - July 27, Cincinnati, OH AIAA 27-5635 Representation of Chemical Kinetics by Artificial Neural Networks for Large Eddy Simulations Baris A. Sen and Suresh Menon School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, GA, 3332-5, USA This paper discusses an approach to incorporate Artificial Neural Network () based kinetics modeling into Large-Eddy Simulation (LES) of reactive flows. The emphasis has been spent for replacing stiff ordinary differential equation (ODE) solvers with for chemical kinetics calculations. A back-propagation type of an code has been developed and its performance has been tested in laminar and turbulent methane/air and syngas/air premixed combustion processes. The is trained using an independent premixed flame calculation and then used as is in the LES. Results indicate that the accuracy of predictions for turbulent flow computations highly depend on the initial training data-set. If well trained, can succesfully predict the chemical state-space using less memory than a conventional look-up table approach and in a computationally efficient manner than the stiff ODE solvers. I. Introduction Accurate prediction of the scalar and the vectoral fields inside the combustion chamber of a gas turbine engine is a challenging task and requires the solution of a three dimensional, highly unsteady, turbulent reactive flow problem. Direct Numerical Simulation (DNS) of the high Reynolds number flows is not affordable yet, since the grid resolution dictated by the solution methodology far more exceeds the current computer capabilities. Therefore, there is a need for simulation models that would decrease the computational cost without significant reduction in accuracy. In the realm of reactive flow computations, great effort has been spent to capture the time dependent processes by LES computations. In order to capture the changes in the species composition state-space due to the chemical reactions a stiff non-linear ODE system must be solved. The size of the equation system is determined by the number of species simulated. Thus, for a detailed reaction mechanism, the chemical kinetics calculations are very time consuming and their application to the engineering problems may be restricted similar to DNS due to the need for computational resources. Relatively smaller chemical kinetics mechanisms obtained by using partial equilibrium and/or steady state assumptions are more widely used than the more detailed mechanisms. However, the reduction process from the detailed to the reduced mechanisms have to be handled carefully in order to yield a match between the range of applicability of the assumptions and the operating conditions of the problem. 2 Remedies for speeding-up chemical kinetics calculations exist in the literature, such as Intrinsic Low- Dimensional Manifolds (ILDM) 2 and In-Situ Adaptive Tabulation (ISAT) methodologies. 3 ILDM requires generation of a multi dimensional look-up table prior to starting the flow computations. The number of resolved states by ILDM should be sufficient enough to include all the possible states that may occur during the simulation. Since this information is not known apriori, huge tables can be generated, which would require large memory for table storage. 4 ISAT, on the other hand, is no different than direct integration of the chemical equation system in the early stages of the simulation, since the look-up table is generated as simulation proceeds. Graduate Research Assistant, AIAA Student Member. Professor, AIAA Associate Fellow. of 5 Copyright 27 by the, Inc. All rights reserved.

Another solution is 5 methodology, which is a relatively new technique in the area of chemical kinetics and its aplicability has not been tested fully yet for turbulent reactive flow applications. is widely used for function approximation, classification, time series prediction, filtering, data association and optimization, 6 and is a computing system made up of a number of simple highly interconnected processing elements (PEs) -neurons-, which process information by its dynamic state response to external inputs. 7 Contrary to the conventional computer modeling approach, which relies on solving the equations, learns by example and they provide predictions into new states. By technique, instead of solving complex governing equations of the system of interest, it is possible to train the PEs using a set of input/output pairs that are representative of the physical process. In the literature there exist several promising attempts of using for chemical kinetics calculations. Choi 8 trained an architecture for ignition delay time calculations where they used temperature, pressure, equivalence ratio and exhaust gas re-circulation as input for their predictions. Later, a well mixed reactor calculation was performed which was coupled with the to obtain the ignition delay time rather than performing direct integration of the chemical kinetics mechanism. It was reported that a speed-up of 6 was obtained by using as opposed to doing direct integration. The main problem associated with computations is the generation of train/test data-set which covers the chemical state space of interest. Several different approaches are introduced in the literature, such as parametrizing the state space with respect to the mixture fraction, 9 or defining the limits of the major representative species of the mixture and assigning random values to them, which would allow to calculate the minor species as well. The first approach was tested on a plug flow reactor calculation where a speed-up of 27 was obtained against the direct integration. The predictions obtained with the latter approach was compared with a tabulated chemistry technique. It is reported that the memory requirement for such an computation requires only 24 kbytes as opposed to 9934 kbytes for tabulated chemistry calculations. Finally an integration of ISAT with was also studied by Chen, where was trained based on an already existing ISAT table. Same approach was later re-visited by Kapoor 2 and was applied to a turbulent premixed flame calculation. Although these were first succesfull attempts of integrating into a turbulent flow computation, the generation of an ISAT table and training based on that requires performing the simulation twice which would increase the overall effort spent on the same simulation. Current paper is seeking for a more robust way of constructing the chemical state space to train the and thus using its unique capability to extract information from this data. II. Mathematical Formulation II.A. LES Modeling An LES approach has been employed with a subgrid mixing and combustion model for the current calculations. For this purpose, the fully compressible, unsteady, multi-species Favre-filtered Navier Stokes equations for continuity, momentum, total energy and species conservation are used as: ρẽ t ρũ i t ρ t + ρũ i x i = + x j [ ρũ i ũ j + pδ ij τ ij + τ sgs ij ] = + [( ρẽ x + p)ũ i + q i ũ j τ ji + H sgs i + σ sgs ij ] = i ρỹk t + x i [ ρỹ k ũ i ρ D k Ỹk x j + φ sgs i,k + θsgs i,k ] = ω k () Here, represents Favre averaging operator, and is calculated for a given quantity f as ( ρf/ ρ), where the overbar stands for spatial filtering using a top-hat filter. The subgrid shear stress (τ sgs ij ), the subgrid viscous work (H sgs i ), the subgrid diffusive flux (θ sgs jm ) arises due to the filtering operation and they need to be modeled. Total energy is given as Ẽ = ẽ + 2ũkũ k + k sgs, where ẽ is the filtered internal energy, k sgs is subgrid scale kinetic energy. The filtered pressure is calculated by the filtered equation of state as p = ρr T, 2 of 5

where R is the mixture gas constant. For momentum and energy transport, the major effect of the small scales is to provide dissipation for the energy cascade. Therefore, an eddy viscosity type subgrid model is suitable for the calculation of subgrid stresses, τ sgs ij and enthalpy flux, H sgs i. The subgrid scale eddy viscosity is calculated as ν t = C ν (k sgs ) 2, where is the grid cut-off scale. An additional transport equation for the subgrid scale kinetic energy k sgs is solved, which is in the form of: ρk sgs t + ( ρũ i k sgs ) = P sgs D sgs + ( ρν t k sgs ) (2) x i x i P r t x i where P sgs = τ sgs ũ i ij x j represents the production term, and D sgs = C ɛ ρ ( k sgs ) 2 3, is the dissipation term. The subgrid scale terms are closed as τ sgs ij = 2 ρν t ( S ij S 3 kk δ ij )+ 2 3 ρksgs δ ij, and H sgs i = ρ νt H P r t x i. Here, P r t =, and the two coefficients, C ν and C ɛ have constant value of.67 and.96, based on an earlier calibration. 3 Also based on earlier studies, the subgrid viscous work, σ sgs ij, is neglected for the current LES. II.B. Subgrid Combustion Modeling The subgrid mass flux, φ sgs jm, the subgrid diffusive flux θsgs jm, and the filtered reaction rate ω k all require closure as well. However, a closure at the resolved scale (as done for momentum and energy transport) is not appropriate since combustion, heat release, volumetric expansion and small-scale turbulent stirring all occur at the small scales, which are not resolved in a conventional LES approach. To address this important physics, in our study a subgrid combustion model based on the Linear Eddy Mixing (LEM) 4 model is used to model combustion occurring within every LES cells (called LEMLES, hereafter). With this approach, there is no need to provide explicit closure for φ sgs jm, θsgs jm, and ω k since a more accurate and exact closure is implemented within the subgrid scales. LEMLES methodology and its ability has been reported extensively elsewhere, 5 7 however, for completeness we give a brief overview of the methodology below. Consider the following exact transport equation for the kth scalar Y k, where there is no LES filtering: ρ Y k t = ρ[ũ i + (u i) R + (u i) S ] Y k (ρy k V i,k ) + ω k (3) x i x i In the above equation, V i,k is the k-th species diffusion velocity that is obtained using Fick s law. To describe the LEMLES approach, the convective velocity is separated into three parts as: ũ i + (u i) R + (u i) S that represent respectively, the LES resolved velocity field, the LES resolved subgrid fluctuation (obtained from k sgs ), and the unresolved subgrid fluctuation. Using this definition, Eq. (3) can be recast into the following two-step numerical form: 7 Y k Y k n t LES = [(ũ i + u i) R ] Y n k (4) x i Y k n+ Y k = t+ tles t ρ [ρ(u i) S Y k n x i + x i (ρy k V i,k ) n ω n k ]dt (5) Here, Eqs. (4) and (5) represent large-scale and small-scale processes, respectively, and t LES is the LES time step. The large-scale step advects the subgrid scalar gradient using a 3D Lagrangian process that ensures strict mass conservation and preserves the small-scale scalar structure. Within each LES cell the subgrid scale processes (the integrand in Eq. (5)) represent respectively, the small-scale turbulent stirring, molecular diffusion and reaction kinetics. The integrand is solved on a D line embedded inside each LES cells with a subgrid resolution that is fine enough to resolve the Kolmogorov scale, η. With such a resolution, both molecular diffusion and reaction rate are closed in an exact sense and this is one of the major strengths of the LEMLES strategy. The D line is aligned in the flame normal or the maximum scalar gradient direction and thus, does not represent any physical Cartesian direction. More specifically, the following reaction-diffusion equation (and an accompanying energy equation) is solved on the D LEM line for species mass fraction: ρ Y k m t s = F s m s (ρy k m V s,k m ) + ω m k (6) 3 of 5

Here t s indicates a local LEM timescale and the superscript m indicates that the subgrid field within each LES supergrid is resolved by N LEM number of LEM cells (i.e., m =, 2,..N LEM ) along the local coordinate s (which is aligned in the direction of the steepest gradient). In the above equation, turbulent stirring is represented by F s m and is implemented using a stochastic re-arrangement events that stir the subgrid scalar fields within each LES cell. The location of stirring event is chosen from a uniform distribution and the frequency of stirring is derived from 3D inertial range scaling laws derived from Kolmogorov s hypothesis as: 8 λ = 54 5 νre [( η ) 5 3 ] C λ 3 [ ( η ) 4 3 ] (7) where, C λ =.67. 9 The eddy size (l) is picked randomly from an eddy size distribution f(l) ranging from to η: f(l) = 5 l 8 3 3 where η = N η 5 3 3 5 η Re 3/4, and Re = u /ν is the subgrid Reynolds number. Volumetric expansion due to heat release is implemented locally within the subgrid by expanding the LEM domain. Although not required, re-gridding is employed to keep the total number of LEM cells constant throughout the simulation to reduce programmatic complexity. Re-gridding can introduce some numerical diffusion that is shown to be negligible in compressible flow simulations using an explicit time step t LES (which is limited by acoustic time scale). Once the subgrid evolution has occurred, large-scale advection (Eq. 4) is implemented by a 3D Lagrangian process that determines the amount of mass to be advected and explictly advecting this amount (and the accompanying subgrid scalar fields) across the LES cell faces. Further details of this process is given elsewhere. 7 Finally, the LES-filtered species mass fractions Ỹk (used in the LES equations) are obtained by Favre averaging the subgrid mass fractions, Y m k : II.C. Chemical Kinetics and N LEM Ỹ k = (/ m= N LEM ρ m ) m= (ρy k ) m (8) For a set of elementary reactions containing an arbitrary number of reactants N, and products M, the net rate at which species X k is produced and/or consumed, the reaction rate, can be calculated by using the following formula: ω k = L ν ki q i (9) i= where L is the number of reactions that species X k is involved. Here, q i is the rate-of-progress variable and calculated as: q i = k fi N k= [X k ] ν M ki kbi k= [X k ] ν ki and ν ki = (ν ki ν ki ). The reaction rate calculation through Eq. (9) provides the link between the chemical kinetics and the subgrid scale evolution of the species that was given in Eq. (6). The rate coefficients of the forward and backward reactions, k f and k b, are functions of temperature. They represent the frequency of molecular collision between molecules and probability that a collision will lead to a chemical reaction. The rate coefficients are expressed in the Arrhenius form as, k f (T ) = B f T n f exp[ E a,f RT ], where E a,f is the activation energy for the forward reaction, B f is the frequency factor, and n f is the temperature exponential. As a result, the chemical kinetics calculations require the solution of a non-linear, coupled system of ODEs. The main problem associated with the solution procedure is the stiffness of the system of equations. Minor species have several order of magnitude different time scales than the major species which prohibits application of standard ODE solvers. 2 Thus, chemical kinetics calculations involves stiff ODE solvers, such as, that lead to a time consuming procedure. () 4 of 5

From a chemical kinetics point of view the reaction rates are functions of the temperature, pressure and species mass fractions and the nonlinear system of equations used to determine the reaction rates can be represented as Yi t = f(t, P, Y, Y 2,..., Y NSP ECI ). From an point of view, it is possible to map inputs (T, P, Y i ) to the reaction rates without requiring any knowledge for the system of equations and considering it as a black box. The calculations that will be presented within the paper relies on this fact and the procedure for training the and its application more specifically will be explained in the following sections. III. Results and Discussion III.A. Code Validation For this particular study a multilayer perceptron type of artificial neural network (MLP-) is used to replace chemical kinetics calculations. A back-propagation 6 algorithm, which is essentially based on the gradient descent procedure, is used for data training. For a back-propagation type of training procedure of MLP-s the algorithm basically consists of two parts: (i) forward propagating the input and (ii) backward propagating the error. Within this context, the output of a single neuron i at an arbitrary layer n M can be calculated as: y n i = g(net n i ), where net n i = W n im y n m, M is the number of neurons connected m= to neuron i, g is the activation function, W n im is the coefficient of the connection between neurons i and m at layers n and n-, and y n m is the output of the neuron m at layer n-. Once the output of each neuron has been found, the calculation of the error associated with each neuron at the output layer is trivial I and given as E i = 2 [d k i y k i ], where d k i is the desired output and y k i is the actual output of neuron i= i at the output layer k. The gradient descent rule involves adjusting the connection weights according to the gradient of the local error across each connection as w = η E, where η is the learning coefficient. Considering the fact that the error function at the output layer is differentiable, with a help of little algebra, the derivative of the local error with respect to the connection coefficient can be obtained for a connection between output and hidden layer as and for the rest of the connections as de dw ij k = [d i k y i k ]g (net i k )y j k () de dw k = [ L δ l w k li ]g (net k k 2 i )y j ij l=i δ l is the local error term of PE l. Its value can be calculated as: δ k l = [d k l y k l ]g (net k l ) for k denoting Z the output layer and as: δ k l = g (net k l ). W k+ zl δ k+ z for the hidden layer. z= In this study a back-propagation type of MLP- code has been written in FORTRAN, which employs the formulation described above. The implementation of the gradient descent theory is slightly different than its original version and has (i) batch update and (ii) momentum term options. 2 The batch update relies on updating the coefficients based on cumulative error function, which is basically sum of the local error of a single PE obtained through certain number of iterations. This allows us to calculate new coefficients after introducing several input output pairs. Momentum term, on the other hand, is used for accelerating the convergence of the by adding a portion of the previous value of the connection coefficient to its new value. The training data is introduced randomly to the net in order to avoid to memorize the specific pattern, and to generalize on the nature of the process. 6 The learning coefficients associated with each hidden layer are held constant for this study. The implementation of a more generic approach, which allows calculation of the learning coefficients for each of the individual PEs dynamically by using the Delta- Bar-Delta 22 or Extended-Delta-Bar-Delta 2 did not lead into a drastic improvement in the training phase, and these options will be re-visited later. (2) 5 of 5

% Error f e.42.22.2.82.62.42.22.2 f(x,y) - -.5 -.5 Y.5.5 - - -.5 -.5 Y X.5.5 (a) Surface plot of the prediction obtained by X.4.2.5 -.5 - Actual Test Data Prediction 2 Prediction.6.5 f(x,y).5.8 % e(x,y).67.46.25.4 -.7 -.38 -.59 -.8 -.2 -.4 -.6 -.8 - -.5 (b) Error surface between the prediction and the actual data Y.5 (c) Effect of the number of training data on predictions Figure. Validation of the code on a simple test function The validity of the training code is first tested on a simple function. This is an important step which proves that the code is working fine and can be used as a tool for chemical kinetics calculations. The function of interest is: f (x, y) = Sin(πx)3.Cos(πy)3. The training is achieved by using two different training files. First training file is constructed by points, in x and in y, where as the second one is for 25 points, with 5 points for each variable. The architecture compromises of 3 hidden layers with 5, and 4 PEs at each layer. This specific distribution is achieved by experimenting over different architectures and checking the performance with the maximum root mean square (RMS) error. The trained nets are used later to predict the function at locations for which it was not trained and the surface that is calculated with is given in Fig. a for the net trained for 25 points. Here, the general topology seem to be accurately captured by the without any significant errors. The absolute error surface, as seen in Fig. b exhibits maximum values corresponding to the regions where the function reaches into its peak. For the rest of the regions, the error seem to be acceptable with a maximum value of.42 %. Fig. c shows the variation of the actual data and predictions on a cross section taken in a x constant plane. As it can be clearly seen in the figure, the net trained for points fails to predict the exact profile especially on the center region where it should be flat. The other net that is trained for 25 points, on the other hand, is very accurate and follows the correct profile in a better way. III.B. Laminar Premixed Flame Computation Applicability of the for substituting the chemical kinetics solver is first tested for a simple laminar -D premixed flame computation. A PSI syngas mixture23 is used as the fuel which includes CO2 and H2 O as diluents in a CO-H2 fuel mixture. The main fuel components exhibit distinct reaction pathways. Hence, combustion simulations for syngas mixtures require relatively large chemical kinetics mechanisms to capture the correct physics,24 which increases the computational cost. can be used to decrease the cost for these calculations. The initial test/train data required for procedure is generated by using PREMIX module of CHEMKIN for an equivalence ratio of.6, and a 4 species, step reduced mechanism is used for representing the chemical kinetics. The mechanism is accurate for a wide range of operating conditions of syngas flames.25 Result obtained from the PREMIX calculation resolved the flame on 256 points. Although an exact definition for the number of train/test data is not given in the literature, it is advised that for an MLP- to generalize properly, the number of independent weights need to be fewer by a factor of than the number of input/output training points.6 Considering the fact that number of PEs used for most of the previous -chemical kinetics calculations is on the order of, 5 points for the current computation is appropriate for the training procedure. Thus, the profiles are populated to 5 equi-distance data points and among these points randomly 7 % is selected for training and 3 % for testing. Similar to the selection 6 of 5

.8.8 Normalized CO Reaction Rate.6.4.2 -.2 -.4 -.6 Test Data Prediction Normalized H Reaction Rate.6.4.2 -.2 -.4 -.6 Test Data Prediction -.8 25 5 75 25 5 Number of Input/Output -.8 25 5 75 25 5 Number of Input/Output (a) CO reaction rate (b) H reaction rate Figure 2. Actual test data and predictions.5e-3.2e-3 Computation Computation.6E+3.4E+3.2E-.E- Computation Computation H Mass Fraction 9.E-4 6.E-4 Temperature [K] H Mass Fraction.2E+3.E+3 8.E+2 CO Mass Fraction Temperature [K] 8.E-2 6.E-2 4.E-2 CO Mass Fraction 3.E-4.E+.25.5.75. Distance [m] 6.E+2 4.E+2 Computation Computation.25.5.75. Distance [m] 2.E-2.E+.25.5.75. Distance [m] (a) H mass fraction (b) Temperature (c) CO mass fraction Figure 3. Application of to a laminar, time dependent syngas-air flame of exact number of training points, the relative percantage of train and test data is also case dependent, and the range varies between 6-8 % in the literature. 8, 9 The architecture for this particular case consists of 5 inputs and 3 hidden layers with each having 2, and 8 PEs, respectively. The optimum value for the architecture is defined by experimenting with different values of PEs at the hidden layers and checking the performance of the for predicting the test data. The architecture used for this study corresponds to an RMS error at the outpur layer of.56 % at maximum for the OH radical. The errors associated with other species are less than this value, e.g., the maximum RMS error for H 2 is.73 %. 4 independent nets are generated for the reaction rates of each species. Following the information given in the literature, 2 in order to avoid saturation at the output layer, the normalized reaction rate is scaled to vary between ±.8 as opposed to the normalized input, which is between ±. The comparison of the predictions and the actual test data is shown in Fig. 2a and b. In general, reaction rates exhibit two distinct profiles depending on the species. Major species, such as CO, H 2, CO 2, H 2 O for this particular case, are either consumed or produced at all in the flame region corresponding to a profile as given in Fig. 2a. The minor species, on the other hand, are consumed and produced in the flame region. A typical example for this kind of behavior is given in Fig. 2b. It is crucial and must be proved that is able to predict both profiles correctly, which is the case for this calculation. For this purpose, the performance of the net has been tested on the data that it was not trained for as shown in Fig. 2a, b with the dashed lines. For both of these species succesfully captures the correct profile, except for small oscillations in H reaction rate right before it starts to decrease. This defect is very small and the RMS value averaged for all the input/output pairs at the at 7 of 5

Normalized Methane reaction Rate.8.7.6.5.4 88 89 9 9 Number of Input/Output (a) Test data and prediction-for all equivalence ratios (b) Test data and predictionclose-up view of 3 equivalence ratios Figure 4. Actual test data and predictions for -MER approach applied to a methane-air flame the output layer is less that %, which is a good indicator of the accuracy of the. In order to further test the performance of the, a time dependent computation is performed by creating an module for chemical kinetics calculations in an existing flow solver. For this purpose, the profiles obtained by PREMIX calculation is used to initialize a 3-D computation. The initial PREMIX profile is linearly interpolated onto a uniform grid of 256 points in the x direction. The same profile is later replicated to all y locations, and the x y plane is copied to all z locations, resulting a grid of 256x2x2 points in x, y and z directions, respectively. The profiles obtained by such a calculation is given in Fig. 3a-c for H, CO mass fraction and temperature. As to yield a better comparison the calculations are also presented in the figures. The H mass fraction exhibits small discrepancey approximately at.5 m. Although not shown here, this defect shows itself for some other minor species too, but major species in general are free from this problem. The error does not impact the resolved flame, as shown in the temperature plot. III.C. III.C.. Flame Turbulence Interaction Methane-Air Flames Computation for a simple steady-state, laminar case is carried out to demonstrate that the developed model for chemical kinetics predictions is working. However, the realistic combustor applications are generally highly turbulent and unsteady problems. LES computations are widely used for the reactive flow analysis and the goal of the current section is to check the applicability of to be used as a chemistry integrator for LES computations. For this purpose, following an earlier LES 26 and DNS 27 study that exist in the literature we consider a premixed methane-air flame front interacting with an initially generated isotropic turbulent flow field. Similar to the previous section, the LES is initialized by using the solution of PREMIX- CHEMKIN and the flame is at a lean condition (φ=.8) with the reactants initially preheated up to 57 K. Based on the initial turbulence level and the flame properties, the premixed flame is in the thin reaction zone (TRZ) regime. A 5 species (CH 4, O 2, CO 2, H 2 O, N 2 ), step reduced methane-air chemical kinetics mechanism is used for training and LES computations. Following the earlier LES, 26 the subgrid scale combustion processes are simulated with LEM by using 2 grid points inside each LES cell. Two different training procedure has been used for this section of the study. First one is based on generation of train/test data-set for a single equivalence ratio (SER), similar to what was done for the laminar computations. The second method, on the other hand, relies on training for multiple equivalence ratios (MER), so as to cover a larger portion of the chemical state space. This is particularly important for turbulence computations with thermally perfect and mixture transport properties, since small perturbations can change the composition space locally. The procedure of training based on a SER is exactly similar to the procedure outlined in the previous section. The solution of PREMIX is populated to 8 of 5

Temperature [K] CH4 Mass Fraction 3 2.4.36.3.25.2.5..5. CH4 Mass Fraction 2.E-2.E-2 SER MER 2 3 4 5 6 (d) CH4 mass fraction profile comparison at cross-section A-A CH4 Mass Fraction.4.36.3.25.2.5..5. Temperature [K] 3 2 2.E+3.5E+3.E+3 (c) CH4 mass fraction surface plot and temperature contours for -MER computation SER MER 4.E-2 3.E-2 2.E+3.5E+3.E+3 (b) CH4 mass fraction surface plot and temperature contours for -SER computation 4.E-2 CH4 Mass Fraction 3 2.4.36.3.25.2.5..5. (a) CH4 mass fraction surface plot and temperature contours for computation.e+ Temperature [K] CH4 Mass Fraction 2.E+3.5E+3.E+3 3.E-2 2.E-2.E-2.E+ 2 3 4 5 6 (e) CH4 mass fraction profile comparison at cross-section B-B (f) Composition state-space accessed by SER, MER training approaches and computations Figure 5. Application of to a turbulent premixed methane-air flame points and several architectures are used to determine which one is better on predicting the test data. Although it is generally advised to employ or 2 hidden layers at most,2 for this particular study 2 hidden layer configuration was the worst case and an architecture with input, output and 3 hidden layers each having 8, 4 and 2 PEs, respectively, is constructed. Based on the second training approach, a database of -D laminar premixed flame solutions for equivalence ratios ranging from.6 to.9 with increments of. is generated. The PREMIX solution for each equivalence ratio compromises 5 points which is interpolated to 5. results based on the test data which is a part of the training procedure is given in Fig. 4a-b. Fig. 4a, shows the predictions and actual test data on the CH 4 reaction rate Temperature hyperplane of the chemical state space. Here, the predictions are comparable with the actual test data and it is not possible to distinguish between both of them for most of the region. The only considerable amount of disrcepancy occurs at the region where reaction rate is at maximum. The reason for this is, this region is accessed only by couple of equivalence ratios, thus, is not well populated. A comparison of the profiles with respect to the equivalence ratios is presented in Fig. 4b, where the agreement is quite accurate. Results obtained by using, -SER and -MER computations after 4 flow through time is introduced in Fig. 5. Here, flow through time is defined as l/u, where l is the largest eddy size and u is the turbulent fluctuation. The effect of different turbulent scales on the scalar field evolution is evident in the 9 of 5

Table. Simulation parameters for flame turbulence interaction problem u /S L l/l f Regime [m] Box Size [cm] FTI-Case 5 TRZ 3. FTI-Case 2 5 5 TRZ 5. Table 2. Flame Parameters used for flame turbulence interaction Fuel Composition Equivalance Ratio S L [m/s] l f [m] H 2 :CO:CO 2 :H 2 O.678:.3:.254:.65.6.25 8.755 4 O 2 :N 2.57:.578 CH 4 mass fraction surface plot obtained by computations as shown in Fig. 5a. The temperature contours covers the maximum value of the CH 4 on a thin region indicating the flame zone. This feature is absent in the -SER computations as given in Fig. 5b. The temperature contours seem to be spread out, covering a larger area across the flame front. Also, the flame topology seems to be very different than the calculations and do not exhibit any similar structure. On the other hand, -MER computations predict a very similar flame topology with calculations as shown in Fig. 5c. A better understanding of the effect of different training approaches can be achieved by investigating the actual profiles along a cross section of the flame and is introduced in Fig. 5d-e. The profiles are obtained across the flame front at y=.25 m and. for Fig. 5d and e, respectively. These locations correspond to a convex and a concave region of the flame. The -MER computation as seen in Fig. 5d, predicts a slightly larger flame region as opposed to. In Fig. 5e, on the other hand, the agreement is better. As it was observed in the surface plot, the predictions obtained by -SER, is very different than the results and look unphyisical. The region on the CH 4 mass fraction and temperature hyperplane of the chemical state-space that is accessed by computations and covered by the MER-SER training procedures is shown in Fig. 5f. Here, the -SER results are shown with black circles and it corresponds to a line on the state-space. The actual data obtained by computations is not a single line, but it covers a region within the state-space. The -SER approach, therefore, can not provide predictions, since the region of interest is outside of its training region. The -MER approach covers a larger space on the state-space which is more general and adapts itself to local perturbations in the composition state space in a better way. III.C.2. Syngas-Air Flames The applicability of into flame-turbulence interaction computations by using LEMLES is studied further in a configuration with an embedded pair of coherent vorticies in an isotropic background turbulence, where there is wrinkling of the flame surface both by large and small scales. More specifically a flame front interacting with a pair of spanwise vorticies superimposed on an isotropic turbulent premixed mixture is of interest. This is more difficult test case as large scale wrinkling alters significantly the diffusion of the species across the flame front. uc, max/s L and D c /l f for this vortex pair is 5 and 5, respectively. The fuel composition and the size of the coherent structure are held constant, and in connection with an earlier study. 28 Two different background turbulence levels are simulated, where the flames are in the thin-reactionzone (TRZ) regime. A 64 3 grid is used for all cases, and the size of the computational domain is selected as to yield a minimum spacing of 3.3 times the Kolmogorov length scale. In the subgrid level within each LES cell, 2 LEM cells (N LEM = 2) are used, which allows a resolution of 4 times smaller than η. The fuel composition is PSI as defined by General Electric. 23 Simulation parameters and flame properties for FTI computations are summarized in Tables and 2, respectively. As it was shown in the previous section, since an -MER approach is superior to -SER, the training approach here uses only the MER technique. The training data is constructed for equivalence ratios ranging from.45 to. with jumps of.5. Also, to avoid poor training due to the few number of 5

.8 Normalized CO Reaction Rate Normalized CO Reaction Rate.8.6.7.4.2.6 -.2 -.4 -.6 -.8 5 5 2 Number of Input/Output.5 25 (a) Test data and prediction-for all equivalence ratios 24 26 28 Number of Input/Output 3 (b) Test data and predictionclose-up view Figure 6. Actual test data and predictions for an -MER approach applied to a syngas-air flame (a) Vorticity surface plot and temperature contours for Case (b) Vorticity surface plot and temperature contours for Case 2 Figure 7. Application of to a turbulent premixed syngas-air flame of data points associated with each single equivalence ratio, the gridding sensivity of PREMIX is increased by adding more grid points to sections where there are large gradients, which yielded approximately 7 points. In total, there is 6728 train and 28768 test data. The architecture consists of 5 layers, having 3 hidden layers with 2, and 8 PEs, respectively. The exact architecture is constructed following the same logic that was outlined in the previous sections. The test data and the predictions for the normalized CO reaction rate is given in Fig. 6a. The data corresponds to an almost non-smooth function of the reaction rate with respect to the equivalence ratio. The can, however, predict the profile fairly well. A close-up look of the profiles reveal that as there are more train/test points for each equivalence ratio, the predictions match with the actual data better than the earlier case. Although it can not be detected well in these figures, the only considerable amount of difference between predictions and the test data is at the regions where there is not any reaction rate which corresponds to a value of.8 on the normalized data. fails to predict an absolute zero value of reaction rate in the post and pre flame regions. This problem is overcome by adding a threshold to the value, beyond which the reaction rates are considered to be zero. The LESLEM results obtained for both cases by employing approach are given in Fig. 7a-b at a non-dimensional time of.3. The time is non-dimensionalized with the characteristic time, which is calculated based on the backgorund turbulence level as Dc /u. Here, Figs. 7a and b shows the contour plot of 5

CO Mass Fraction..8.6.4.2. 2 3 4 5 6.3E-3..E-3.8.6.4 H2O Mass Fraction O Mass Fraction 7.5E-4 5.E-4.2 2.5E-4..E+ 2 3 4 5 6 OH Mass Fraction.5E-3.E-3 5.E-4.E+ 2 3 4 5 6 (a) CO and H 2 O mass fraction for Case (b) O mass fraction for Case (c) OH mass fraction for Case Figure 8. Comparison of species profiles obtained by and approaches for turbulent, premixed syngas-air flame, Case CO Mass Fraction..8.6.4. 3.E-3.8.6.4 H2O Mass Fraction O Mass Fraction 2.E-3.E-3 OH Mass Fraction 3.E-3 2.5E-3 2.E-3.5E-3.E-3.2. 2 3 4 5 6.2..E+ 2 3 4 5 6 5.E-4.E+ 2 3 4 5 6 (a) CO and H 2 O mass fraction for Case 2 (b) O mass fraction for Case 2 (c) OH mass fraction for Case 2 Figure 9. Comparison of species profiles obtained by and approaches for turbulent, premixed syngas-air flame, Case 2 of temperature superimposed on the vorticity surface plot. The flame front is wrinkled both by the large scale-coherent vortex pair and the small scale background isotropic turbulence. The small-scale vorticity in the incoming turbulent field dissipates across the flame and only the large structures exist in the burned side due to the high temperature values in this region. A detailed description of the interaction between the flame and the vortical structures is given elsewhere. 28 The comparison of the computed profiles across the flame is given in Fig. 8a-c for Case. The profiles obtained for major species CO and H 2 O match well with the computations. The minor species, O and OH, show discrepancy near the peak region. The maximum error for both of the species is less than 8 %. However, both species mass fraction is on the order of 3 and the impact of their error is not very severe on the major features of the flame. The agreement between the profiles obtained for flame Case 2, as shown in Fig. 9a-c, however, is not as good as it is obtained for Case. CO starts to be consumed earlier than that is calculated with and the H 2 O mass fraction exhibits a peak value within the flame zone. The maximum value of the minor species is overpredicted, almost 4 times for O and 2 times for OH. The profile for OH also corresponds to an oscillation which is not physical. In order to gain further insight into the effect of on the LEMLES computations, the CO 2 OH hyperplane of the chemical state-space is presented in Fig.. Here, Fig. a corresponds to the region that was accessed by, whereas Fig. b is by the calculations. In 2 of 5

(a) Actual sample space accessed by computation (b) Training data for and the actual sample space accessed by computations (c) Refined training data for and the actual sample space accessed by computations Figure. CO 2 OH Hyperplane Accessed during computations and training Phase O Mass Fraction 3.E-3 2.E-3.E-3 refine OH Mass Fraction 3.E-3 2.5E-3 2.E-3.5E-3.E-3 refine 5.E-4.E+ 2 3 4 5 6.E+ 2 3 4 5 6 (a) O mass fraction profiles by and computations for Case 2 (b) OH mass fraction profiles by and computations for Case 2 Figure. Comparison of species profiles obtained by and approaches for turbulent, premixed syngas-air flame, Case 2 with refined table Fig. b, the region that was covered within the initial training phase is also given, and denoted with black dots. The computations, similar to that was obtained for earlier methane-air case, supports the idea why it is not possible to achieve training and expect accurate results for a SER computation. This time, rhe region is more complex and can not be represented by a SER, which would actually correspond to a single line on the hyper-plane. A direct comparison of the region covered by the training phase and the actual region constructed by computations reveal that only a single portion of the state-space is actually accessed. Since the is trained for a wide range, the accessed points with the computation is spread over the entire region. Moreover, the training table is sparse in most of the region, which would avoid to fully generalize on the training data. In order to get better solution, a new training table is produced for a smaller portion of the chemical state-space but with a finer resolution. The is trained for a range of equivalence ratios from.5 to.7 with. increments. The CO 2 OH hyperplane representative of the region accessed through training and LEMLES computations is given in Fig. c. The accessed region by LEMLES for this training file is very similar to the region covered by, as shown in Fig. a. For low values of the CO 2, the region used for training becomes very thin and cannot cover the accessible region as predicted by 3 of 5

calculations. Hence, the prediction spreads over a relatively larger area compared to. The species profile as seen in Fig. 9. is much better than the results in Fig.. There is still an overestimation of the minor species maximum value, but the error is smaller than the previous case and most importantly the profiles do not exhibit any non-physical structure. Nevertheless this is going to be studied further for future study. As it was outlined in the previous sections, most important merit of using against stiff ODE solvers is its speed, and against ISAT and ILDM based look-up table approaches is its efficiency in terms of memory. The time required for one time step iteration for the laminar, premixed syngas-air flame simulation with is.2547 sec and for is.627, which corresponds to a speed up of 3.47. The size of the table that is used for the training procedure is 23.8 MB. For the turbulent syngas-air flame simulation, the time required for one iteration is 42.7969 sec for and.62 sec for, which approximately corresponds to the same amount of speed up, 3.68. For the turbulent flame simulation, compared to the laminar flame, the memory required for the training is larger since a MER approach is employed. The table that is used for training is 522.6 MB and hence, using saved a considerable amount of memory. Finally, for the turbulent methane-air flame a speed up of.98 is achieved. The reason why the amount of speed up decreased compared to the syngas-air flames is related with the reduced mechanism that is used. The chemical kinetics calculations for the methane-air calculations are made by using a small, step 5 species reduced mechanism. Thus, the time required for the solution of the stiff ODE is not as huge as it is for the syngas flame. Still, saves a 93.4 MB of memory. IV. Conclusion Calculation of the changes in the chemical state-space due to the reactions, relies on employing stiff ODE solvers. This approach may be very time consuming if the chemical kinetics mechanism used for the simulations involves many species and steps. Look-up table approach, on the other hand, can be used for decreasing the cost, but the memory requirement for this approach is usually huge. The current paper seek for an optimal solution into this problem by using approach. It has been showed that, in contrast to employing stiff ODE solvers, is very time effective, and the memory requirement is not as large as it is for look-up table approach. The focus of the paper has been spent for the determinig the reliability of the predictions, which is highly dependent on the training procedure. For premixed flames it has been showed that the can be trained based on the solution of a D laminar premixed flame. This approach, however, is not very effective for turbulent computations, since turbulence may perturb the composition state-space locally which would not let it to be represented simply by one equivalence ratio. Training based on multiple equivalence approach yielded better results as it has been shown for turbulent, premixed methane-air and syngas-air flames. As long as the composition state-space is covered with sufficient number of chemical states, can be used succesfully for replacing stiff ODE solvers. As for the next step of incorporating into LES computations for decreasing the computational time without sacrificing the accuracy is to use it as a subgrid scale combustion model. This would involve more complicated architecture but would drastically decrease the computational cost. The study is still underway. Acknowledgements The authors would like to acknowledge Prof. J.-Y. Chen for providing the reduced chemical kinetics reaction mechanism for syngas that was used in this study. Appendix IV.A. Step reduced Chemical Kinetics Model Species:H 2, H, O, OH, HO 2, H 2 O, CO, O 2, H 2 O 2, CO 2, CH 4, CH 2 O, NO, N 2 H + O 2 = O + OH H 2 + O = H + OH H 2 + OH = H + H 2 O 4 of 5

2H = H 2 H + O 2 = HO 2 2OH = H 2 O 2 O + CO = CO 2 O + CH 4 = 2H + CH 2 O O + O 2 + CH 2 O = OH + HO 2 + CO References Goldin, G. and Menon, S., Scalar PDF Construction Model for Turbulent Non-Premixed Combustion, Combust. Sci. Tech., Vol. 25, 997, pp. 47 72. 2 Maas, U. and Pope, S. B., Simplifying Chemical Kinetics: Intrinsic Low-Dimensional Manifolds in the Composition Space, Combust. Flame, Vol. 88, 992, pp. 239 264. 3 Pope, S. B., Computationally Efficient Implementation of Combustion Chemistry Using In Situ Adaptive Tabulation, Combust. Theory Modelling, Vol., 997, pp. 4 63. 4 Sen, B. A., Ozdemir, I. B., and Warnatz, J., Implementation of the Hierarchical Structures of the Combustion Processes into Intrinsic Low-Dimensional Manifolds Technique, 4th International Symposium on Turbulence, Heat and Mass Transfer, 23. 5 Mehrotra, K., Mohan, C. K., and Ranka, S., Elements of Artificial Neural Networks, MIT Press, 997. 6 Christodoulou, C. and Georgiopoulos, M., Applications of Neural Networks in Electromagnetics, Artech House, 2. 7 Hecht-Nielsen, R., Neurocomputing, Addison-Wesley, 99. 8 Choi, Y. and Chen, J.-Y., Fast Prediction of Start-of-Combustion in HCCI with Combined Artificial Neural Networks and Ignition Delay Model, Proc. Combust. Inst., Vol. 3, 25, pp. 27 278. 9 Blasco, J. A., Fueyo, N., C., L. J., Dopazo, C., and Chen, J.-Y., A Single-Step Time-Integrator of a Methane-Air Chemical System Using Artificial Neural Networks, Computers and Chemical Engineering, Vol. 23, 999, pp. 27 33. Blasco, J. A., Fueyo, N., Dopazo, C., and Ballester, J., Modelling the Temporal Evolution of a reduced Combustion Chemical System With an Artificial Neural Networks, Combust. Flame, Vol. 23, 998, pp. 38 52. Chen, J. Y., Blasco, J. A., Fueyo, N., and Dopazo, C., An Economical Strategy for Storage of Chemical Kinetics:Fitting InSitu Adaptive Tabulation With Artificial Neural Networks, Proc. Combust. Inst., Vol. 28, 2, pp. 5 2. 2 Kapoor, R., Lentati, A., and Menon, S., Simulations of Methane-Air Flames Using ISAT and, AIAA--3847, 2. 3 Chakravarthy, V. and Menon, S., Large-Eddy Simulations of Turbulent Premixed Flames in the Flamelet Regime, Combust. Sci and Tech, Vol. 62, 2, pp. 48. 4 Kerstein, A. R., Linear-Eddy Model of turbulent Scalar Transport and Mixing, Combust. Sci. Tech., Vol. 6, 988, pp. 39 42. 5 Menon, S., Yeung, P.-K., and Kim, W.-W., Effect of Subgrid Models on the Computed Interscale Energy Transfer in Isotropic Turbulence, Computers and Fluids, Vol. 25, No. 2, 996, pp. 65 8. 6 Eggenspieler, G. and Menon, S., LES of Premixed Combustion and Pollutant Emission in a DOE-HAT Combustor, AIAA-23-39, 23. 7 Menon, S. and Patel, N., Subgrid Modeling for LES of Spray Combustion in Large-Scale Combustors, AIAA Journal, Vol. 44, No. 4, 26, pp. 79 723. 8 Kerstein, A. R., Linear-Eddy Model of Turbulent Transport II, Combust. Flame, Vol. 75, 989, pp. 397 43. 9 Chakravarthy, V. and Menon, S., Linear-Eddy Simulations of Reynolds and Schmidt Number Dependencies in Turbulent Scalar Mixing, Physics of Fluids, Vol. 3, 2, pp. 488 499. 2 Warnatz, J., Maas, U., and Dibble, R. W., Combustion:Physical and Chemical Fundamentals, Modeling and Simulation, Experiments, Pollutant Formation, Springer, 26. 2 NeuralWare, Neural Computing. A Technology Handbook for NeuralWorks Pro fessional II / PLUS, NeuralWare, 2. 22 Jacobs, R. A., Increased Rates of Convergence Through Learning Rate Adaptation, Neural Networks, Vol., 988, pp. 295 37. 23 Bradar, R. D. and Jones, R. M., GE IGCC Technology and Experience with Advanced Gas Turbines, Ger-427, GE Power Systems, Schenectady, NY, October 2. 24 Natarajan, J., Nandula, S., Lieuwen, T., and Seitzman, J., Laminar Flame Speeds of Synthetic Gas Fuel Mixtures, ASME Paper, 25. 25 Mosbacher, M., Haynes, J., Janssen, J., Brumberg, J., Lieuwen, T., Menon, S., Seitzman, S., and Anand, A., Fuel- Flexible Combustion System for Co-Production Plant Applications, De-fc26-3nt4776, GE Global Research, Schenectady, NY, October 26. 26 Sankaran, V. and Menon, S., Subgrid Combustion Modeling of 3-D Premixed Flames in the Thin-Reaction-Zone Regime, Proc. Combust. Inst., Vol. 3, No., 25, pp. 575 582. 27 Trouvé, A. and Poinsot, T., The Evolution Equation for the Flame Surface Density in Turbulent Premixed Combustion, Journal of Fluid Mech., Vol. 278, 994, pp. 3. 28 Sen, B. A. and Menon, S., Large-Eddy Simulations of Turbulent Flames in Syn-Gas Fuel-Air Mixtures, AIAA-27-435, 27. 5 of 5